Artificial Intelligence in Healthcare: Hope or Hype?

by Ray McBeth, PhD

Forms of Artificial Intelligence (AI) have been around since the 1950s. Until recently, development has been slow, and the technology has not been altogether reliable and also very expensive. But advances in processing capability, analytics, and big data have changed much of that. We take it for granted that our smart phone can accurately hear us when we ask about a restaurant and then provide us with step-by-step driving directions to that restaurant. Soon, some predict that our “smart” car will actually drive us there with no involvement on our part at all. These abilities are all due to various systems using AI.

So what is Artificial Intelligence?

There are actually a number of different definitions and sub-definitions. In simplified terms, AI follows either a “structured” or “unstructured” approach. Structured approaches start with data that has already been categorized by humans (e.g., pictures of dogs or cats). Often millions of images are reviewed and the system is given feedback on the accuracy of its analysis until the system has a sense of what makes dog a dog (or not a dog) or what makes a cat at cat (or not a cat). The system is then able to make accurate identification choices when shown new pictures of dogs or cats and then integrate that new learning into its overall understanding. Unstructured approaches, on the other hand, start with large amounts of data that have not been categorized. The system then begins to review it to see what patterns exist. There are so many variables involved that no person or group of people could possibly keep them all straight. There are not yet any truly fully unstructured systems, but there are some semi-unstructured algorithms in development.   

While AI is becoming more and more common in our everyday lives, it is only beginning to have an impact on healthcare. One might ask, why is that?

There are a number of reasons and some case studies can help illustrate the answer.

One of the promoted uses of AI in healthcare is in diagnostic imaging (e.g., interpretation of x-rays); this is an example of a structured approach. For example, machines can scan chest x-rays to see if a lung is normal or abnormal (e.g., collapsed). They can do this much faster and more accurately than even the most seasoned radiologist. However, in order to do this, the machine has to view literally millions of lung images that have already been verified by a human as either normal or abnormal. The goal is not to replace the radiologist, but to free the radiologist from verifying “normal” images (in itself a somewhat boring and “mindless” process) so that s/he is available to spend more time interpreting those images that are not normal. Similar approaches can be applied to other conditions identified by imaging technology or other diagnostic criteria. This will become more and more important for two reasons; older radiologists and other diagnosticians are retiring and fewer are graduating from medical school; in addition there are many geographic areas where very few specialists are even available.  

This leads to another question . . . If the technology is that good, why is it not in greater use?

First of all, the research that is conducted to have the machine learn to interpret those x-rays is very expensive and often machine dependent. There are currently few payment models that will allow the hospital to get paid for a machine to read an x-ray and judge it to be normal or not and then refer the ones identified as not normal to a radiologist for further review.

Then there is the issue of trust. Much of AI has a sort of “black box” feel, meaning practitioners don’t trust what the machine “learned” because they can’t see how it happened. This causes them to then doubt the results. A similar outcome occurred when business processes, like accounting, were first automated.

There is also a basic competency concern. For example, if radiologists only read abnormal x-rays, will their understanding of what is normal become degraded? Or will they begin to assume that the AI is always accurate and not take then next step? This is similar to the concern regarding large commercial aircrafts which virtually fly themselves and the discovery that, because of that, pilots are losing their skills. Airlines have, in fact, sent letters to pilots encouraging them to actually fly their planes more to overcome this.

Finally, there are many unresolved legal, ethical and regulatory concerns about safety, privacy, liability, and the like. These are holding AI back from being integrated especially into the clinical aspects of the healthcare industry.

The x-ray example is a fairly simple one, but consider if other imaging data (e.g., CT, MRI) along with past health history (including lab results), genomic data, and lifestyle data were also able to be integrated and analyzed to identify patterns (a semi-unstructured approach). What might be possible, then?

Some are suggesting that precision medicine will occur. That is, ongoing prediction and treatment processes designed specifically for each individual. In identifying patterns, the algorithm would take into account gender, age, past health concerns, social status, geographic location and other information that might otherwise be viewed as irrelevant or unmanageable. Then the algorithm will continue monitoring the effects of various interventions to improve outcomes.  This is actually beginning to happen as more data is integrated into what is known about each of us. As suggested earlier, this is something that only a machine can do; there is simply too much data, too many images, and too many test results for any one human to keep track of and the volume of information on each of us is growing daily. We are a long way from full implementation of that type of medicine, but it is beginning.    

While the clinical uses of AI are likely to only slowly be adopted, there are many potential uses for AI in healthcare. A few of them include:

  • Automation, fact checking, error correction, and analysis of many administrative functions including supply chain, medical billing, accounting, and other routine tasks (1)

  • Optimizing scheduling of patients, staff, and facilities (2)

  • Reduction in adverse health effects including diagnostic considerations, gaps in care, hospital-based infections, medication errors, post-discharge complications, unnecessary readmissions, and the like (3)

  • Identification of the most successful combination of treatment interventions (e.g., quickest recovery time, fewest fatalities) for a given type of surgery both in general and based on individual characteristics (e.g., age, gender) (4)

  • Ongoing monitoring and reporting of existing or potential health conditions based on each individual's specific needs, through the Internet of Things (5)

  • Reviewing research, literature, and/or current best practices to provide clinicians with the most current evidence-based interventions in their fields (6)

  • Use by insurers (since they have the most complete picture of all of the health history of any patient) being able to manage risk and propose interventions including those associated with the social determinants of health (7)

  • Use of chatbots to gather data and interact with patients and robots to deliver medicine and supplies (8,9)

  • Use by patients to determine the most cost effective insurance based on their current and predicted health needs (10)

  • Evaluation of the efficacy of various compounds (i.e., drug testing) in the treatment of illnesses (11)

The initial uses of AI are likely to be in areas where it replaces repetitive activity. This will lead to a loss of both blue and white collar jobs, but it is also likely to lead to new types of work. These new types of work may well be an extension of what was replaced, but with a distinctly human focus. Every new technology has replaced existing jobs, but has inevitably replaced them with new types of work. Consider the many new job titles (e.g., data scientist) that didn’t even exist even a few years ago.

Even in clinical areas it is important to note that while machines have become very good at pattern recognition and will become even better as they process more similar images, that is still not understanding. Specialists will always (at least for the foreseeable future) be needed to supply that understanding, but with the support of AI they will be able to do so more quickly and more accurately.

While there is a lot “hype” about AI in healthcare (mostly due to heightened expectations that have not yet been met) there is hope that in the future, AI will enhance the ability of healthcare providers to spend more time with patients doing what only they can do and doing it in a way that provides better and more cost effective care.

(1) AI-fueled Organizations, Nitin MittalDavid KuderSamir Hans

 (2) Big Data, Analytics & Artificial Intelligence: The Future of Health Care is Here, General Electric Company, 2016,%2520Analytics%2520&%2520Artificial%2520Intelligence:%2520The%2520Future%2520of%2520Health%2520Care%2520is%2520Here

(3) Technology Trends in Healthcare and Medicine: Will 2019 be Different? Dr. Chris Stout

(4) How a Tiny Community Hospital Deployed AI to Reduce Costs and Enhance Patient Outcomes;F:US!100&PreviousLoginCount=14&ForceProfileToBeFilledOut=0&DisplayItem=EH303214&ShowKey=32883&ShowFrameFormatOverride=NULL&RandomValue=1551029585780

(5) IoT Analytics for Health, SAS

(6) How Conversational AI Can Help Cure Physician Burnout

(7) How AI can cut health care costs

(8) Vanderbilt U rolls out AI voice assistant for EHRs

 (8,9) Medical Robots That Could Change Healthcare

 (10) How AI is transforming healthcare and solving problems in 2017

 (11) Pharma researchers turn to AI for help in search for new drugs

Brea Neri